For training graphs (when inputs require grad), previously, we would speculate the forward and backward graph to determine if there are any graph breaks, side effect and etc but would not actually use these speculated graphs. We would just insert a call function node on the graph and later rely on autograd's tracing.
This approach does not work for more generalized graphs like graphs that include user defined triton kernels because autograd is not able to do the higher order function conversation.
This PR speculates the forward and backward functions and emits them in a HOF that later gets used via templating mechanism.
While working on this PR, I have exposed some bugs in the current tracing due to trampoline functions losing the source information resulting in incorrect graphs being produced. I have fixed these source information bugs and killed the trampolines.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/116358
Approved by: https://github.com/jansel
Recent 2 triton PRs (https://github.com/openai/triton/pull/2701, https://github.com/openai/triton/pull/2756) change the interface for triton.compile, this PR added the necessary change on inductor side to work with both old and new compile API.
Also there is some simplification between compilation call in subprocess and the one in main process
- previously we pass warm_cache_only=True if the compilation happens in subprocess. But triton never use that argument in the currently used pin. So I removed that
- previously we only pass compute_capability if compilation happens in subprocess. The PR change that to always passing compute_capability to triton.compile no matter if the compilation happens in main or sub process.
Updated:
There are more interface change from triton side. E.g.
- tl.math.{min, max} now requires a propagate_nan argument
- JITFunction.run now requires a warmup argument. This affect the benchmarking phase of matmul max-autotune; on the other hand, JITFunction.run forbids stream argument now. Simply removing passing this in when benchmarking matmul triton kernel will work for both old and new version of triton.
- triton Autotuner change attribute name from 'warmup' to 'num_warmup' and from 'rep' to 'num_rep'. This cause dynamo failed to handle triton Autotuner object since dynamo TritonKernelVariable makes assumption about attribute names. It's used in some test cases that a model call triton Autotuner directly.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/115878
Approved by: https://github.com/jansel
Recent 2 triton PRs (https://github.com/openai/triton/pull/2701, https://github.com/openai/triton/pull/2756) change the interface for triton.compile, this PR added the necessary change on inductor side to work with both old and new compile API.
Also there is some simplification between compilation call in subprocess and the one in main process
- previously we pass warm_cache_only=True if the compilation happens in subprocess. But triton never use that argument in the currently used pin. So I removed that
- previously we only pass compute_capability if compilation happens in subprocess. The PR change that to always passing compute_capability to triton.compile no matter if the compilation happens in main or sub process.
Updated:
There are more interface change from triton side. E.g.
- tl.math.{min, max} now requires a propagate_nan argument
- JITFunction.run now requires a warmup argument. This affect the benchmarking phase of matmul max-autotune; on the other hand, JITFunction.run forbids stream argument now. Simply removing passing this in when benchmarking matmul triton kernel will work for both old and new version of triton.
- triton Autotuner change attribute name from 'warmup' to 'num_warmup' and from 'rep' to 'num_rep'. This cause dynamo failed to handle triton Autotuner object since dynamo TritonKernelVariable makes assumption about attribute names. It's used in some test cases that a model call triton Autotuner directly.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/115878
Approved by: https://github.com/jansel
I used a couple of type-ignore comments in ir.py because it constructs
short-lived instances of FixedLayout and GraphModuleSerializer, just to
call a single method on them that doesn't use all their members. Making
those unused members optional would make the rest of the code a lot
messier with sprinkled `assert` statements.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/113534
Approved by: https://github.com/albanD
ATT, there are cases where multiple kernel invocations have same kernel names, and key_averages() will wrongly get average results across different invocations. This fix uses cuda_time_total / n_repeat instead.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/113611
Approved by: https://github.com/chenyang78
We spend somewhere on the order 1% in `sympy.Expr.free_symbols` as it is called millions of times.
Most of the time we actually just want to know "is this a constant", however `e.is_constant()` is
horribly slow. It turns out though that there is another propery `is_number` that does what we want.
> property is_number:
>
> Returns True if self has no free symbols and no undefined functions (AppliedUndef, to be precise). It will be faster
> than if not self.free_symbols, however, since is_number will fail as soon as it hits a free symbol or undefined
> function.
Even further, we also avoid the overhead of building the unnecessary set object.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/112688
Approved by: https://github.com/lezcano
This was originally @jansel's PR:
https://github.com/pytorch/pytorch/pull/102625, which I've built upon.
This diff implements static memory planning. It's disabled by default
while we examine its performance.
We use a greedy-by-size approach. For dynamic shapes, the sizes of the
example inputs are used as estimates when making planning decisions. We
generate expressions to calculate the actual memory offsets and sizes at
runtime when the values of the dynamic shapes are known. In order to
simplify these calculations, we have organized the allocations into a
tree that branches on space (address offsets) and time (live ranges).
Finally, we need to align these offsets, so we have added an `align`
sympy Expr to express these calculations.
Some limitations:
1. It is only enabled during inference for now. Enabling it for training
increases peak memory usage as we allocate all the memory needed for
training upfront, before freeing the memory allocated during
inference. We can probably address this by doing planning for both
the inference and training passes together.
2. It doesn't work with PyTorch Distributed, because kernels like
AllGatherIntoTensor codegen strings which do memory operations. We
can fix this down the line by having them emit MemoryPlanningLines
instead.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/112178
Approved by: https://github.com/desertfire, https://github.com/jansel
This was originally @jansel's PR:
https://github.com/pytorch/pytorch/pull/102625, which I've built upon.
This diff implements static memory planning. It's disabled by default
while we examine its performance.
We use a greedy-by-size approach. For dynamic shapes, the sizes of the
example inputs are used as estimates when making planning decisions. We
generate expressions to calculate the actual memory offsets and sizes at
runtime when the values of the dynamic shapes are known. In order to
simplify these calculations, we have organized the allocations into a
tree that branches on space (address offsets) and time (live ranges).
Finally, we need to align these offsets, so we have added an `align`
sympy Expr to express these calculations.
Some limitations:
1. It is only enabled during inference for now. Enabling it for training
increases peak memory usage as we allocate all the memory needed for
training upfront, before freeing the memory allocated during
inference. We can probably address this by doing planning for both
the inference and training passes together.
2. It doesn't work with PyTorch Distributed, because kernels like
AllGatherIntoTensor codegen strings which do memory operations. We
can fix this down the line by having them emit MemoryPlanningLines
instead.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/112178
Approved by: https://github.com/desertfire, https://github.com/jansel
This is a small fix in "do_bench_using_profiling()".
When CUDA kernels are executed in a non-default CUDA stream, if cuda.synchronize() is called, a CUDA kernel named "Context Sync" will be launched to the default stream to wait until all other streams are finished. This CUDA kernel has "CUDA time" but is not a real kernel to profile. This fix excludes "Context Sync" when calculating kernel total time.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/112223
Approved by: https://github.com/int3, https://github.com/chenyang78
Updates `_export.aot_compile` to pass a torch IR graph to inductor, allowing inductor to now run the pre_grad_passes, and reuse more of inductor's code.
Also updates the API to only return the `so_path`, and not returning the exported program. The pytree call spec is now serialized and placed inside of the generated model code. When calling the model, because there is no c++ pytree implementation linked yet, we can access the call specs through `get_call_spec()`, and call pytree flatten/unflattenin python.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/110020
Approved by: https://github.com/desertfire
Improves perf of llama_v2 locally from 1.55 -> 1.57
The initial heuristic is to lower to pointwise if # of inputs is <= 4, and all the inputs are pointwise or cannot be memory planned away, or if all the outputs are pointwise.
Perf run was +3% on inference.. There are definitely instances where we should be lowering to foreach_kernels, but it's less flexible for fusion. The motivating example was:
```
def rotate_half(x):
"""Rotates half the hidden dims of the input."""
x1 = x[..., : x.shape[-1] // 2]
x2 = x[..., x.shape[-1] // 2 :]
return torch.cat((-x2, x1), dim=-1)
def apply_rotary_pos_emb(q, k, cos, sin):
iota = torch.ops.prims.iota.default(512, start = 0, step = 1, dtype = torch.int64, device = device(type='cuda', index=0), requires_grad = False)
# File: /scratch/eellison/work/torchdynamo/lib/python3.8/site-packages/transformers/models/llama/modeling_llama.py:657, code: position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
unsqueeze = torch.ops.aten.unsqueeze.default(iota, 0)
position_ids = torch.ops.aten.reshape.default(unsqueeze, [-1, 512]); unsqueeze = None
# The first two dimensions of cos and sin are always 1, so we can `squeeze` them.
cos = cos.squeeze(1).squeeze(0) # [seq_len, dim]
sin = sin.squeeze(1).squeeze(0) # [seq_len, dim]
cos = cos[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim]
sin = sin[position_ids].unsqueeze(1) # [bs, 1, seq_len, dim]
q_embed = (q * cos) + (rotate_half(q) * sin)
k_embed = (k * cos) + (rotate_half(k) * sin)
return q_embed, k_embed
```
Also not sure if I should be more worried about concatting reduction->pointwise inputs.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/111233
Approved by: https://github.com/Chillee
Did some easy fixes from enabling TRY200. Most of these seem like oversights instead of intentional. The proper way to silence intentional errors is with `from None` to note that you thought about whether it should contain the cause and decided against it.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/111496
Approved by: https://github.com/malfet
This PR implements intra-graph communication reordering pass on Inductor scheduler IR, based on Horace's previous PR #100762.
Main algorithm:
1. Greedily moves waits as late as possible (i.e. until we reach a use)
2. Greedily moves comms as early as possible (i.e. until we reach an input)
3. Move computes following simple heuristics to improve overlap.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/108091
Approved by: https://github.com/Chillee, https://github.com/wanchaol
Fixes#111066#111065#111064
Currently use_cutlass_template is returning True on ROCm but the feature is not supported. Fix to return false on ROCm. I considering adding this change to `try_import_cutlass` instead but the comments hinted that this function would be removed at some point.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/111132
Approved by: https://github.com/jansel
Summary:
Previously, we link against cuda libs even for pure cpp backend.
This caused issues for cases where the inference platform does not
have GPUs. This diff removed cuda dependency for cpp backend.
Reviewed By: bertmaher, muchulee8, mikekgfb
Differential Revision: D49800712
Pull Request resolved: https://github.com/pytorch/pytorch/pull/110409
Approved by: https://github.com/bertmaher, https://github.com/desertfire
Example of when the `evict_first` heuristic helps.
```
@torch.compile
def f(a, b):
return (a * b).sum(dim=-1)
N = 512
inps = (torch.randn(N, N, N).permute(2, 1, 0), torch.randn(N, N, N).permute(1, 2, 0))
from torch._inductor.utils import do_bench
print(do_bench(lambda: f(*inps)))
```
This generates code like this: http://ix.io/4HFs
```
Original: 3.8 ms
This PR: 3.54 ms
Always `evict_first: 5.4ms
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/108841
Approved by: https://github.com/lezcano, https://github.com/jansel
A resubmit of https://github.com/pytorch/pytorch/pull/108447. Copy over the descriptions:
This is a follow-up of the discussion in https://github.com/pytorch/pytorch/pull/108356, where we want to repalce source_fn with source_fn_stack
Before this PR, for the following example:
```python
backend = EagerAndRecordGraphs()
@torch.compile(backend=backend, fullgraph=True)
def cond_f(pred, pred2, x, y):
def true_fn(pred2, x, y):
return x + y
def false_fn(pred2, x, y):
def true_fn2(x, y):
return x.sin() - y.cos()
def false_fn2(x, y):
return x.cos() - y.sin()
return control_flow.cond(pred2, true_fn2, false_fn2, (x, y))
return control_flow.cond(pred, true_fn, false_fn, (pred2, x, y))
```
The graph captured is shown below:
```python
class GraphModule(torch.nn.Module):
def forward(self, L_pred_ : torch.Tensor, L_pred2_ : torch.Tensor, L_x_ : torch.Tensor, L_y_ : torch.Tensor):
l_pred_ = L_pred_
l_pred2_ = L_pred2_
l_x_ = L_x_
l_y_ = L_y_
cond_true_1 = self.cond_true_1
cond_false_1 = self.cond_false_1
cond = torch.ops.higher_order.cond(l_pred_, cond_true_1, cond_false_1, [l_pred2_, l_x_, l_y_]); l_pred_ = cond_true_1 = cond_false_1 = l_pred2_ = l_x_ = l_y_ = None
return (cond,)
class GraphModule(torch.nn.Module):
def forward(self, l_pred2_, l_x_, l_y_):
add = l_x_ + l_y_; l_x_ = l_y_ = None
return add
class GraphModule(torch.nn.Module):
def forward(self, l_pred2_, l_x_, l_y_):
cond_true_0 = self.cond_true_0
cond_false_0 = self.cond_false_0
cond = torch.ops.higher_order.cond(l_pred2_, cond_true_0, cond_false_0, [l_x_, l_y_]); l_pred2_ = cond_true_0 = cond_false_0 = l_x_ = l_y_ = None
return cond
class GraphModule(torch.nn.Module):
def forward(self, l_x_, l_y_):
sin = l_x_.sin(); l_x_ = None
cos = l_y_.cos(); l_y_ = None
sub = sin - cos; sin = cos = None
return sub
class GraphModule(torch.nn.Module):
def forward(self, l_x_, l_y_):
cos = l_x_.cos(); l_x_ = None
sin = l_y_.sin(); l_y_ = None
sub = cos - sin; cos = sin = None
return sub
```
the source_fn for inner cond, sin, cos will be a (name, target) tuple:
```
('cond', <torch._ops.HigherOrderOperator object at xxx>)
('sin', 'sin')
('cos', 'cos')
('sub'. <built-in function sub>)
```
After this pr, the source_fn_stack will be a list of (name, target) tuple. The bottom of stack is the end of the list.
```
[('cond', <torch._ops.HigherOrderOperator object at xxx>), ('cond', <torch._ops.HigherOrderOperator object at xxx>)],
[('cond', <torch._ops.HigherOrderOperator object at xxx>), ('cond', <torch._ops.HigherOrderOperator object at xxx>), ('sin', 'sin')],
[('cond', <torch._ops.HigherOrderOperator object at xxx>), ('cond', <torch._ops.HigherOrderOperator object at xxx>), ('cos', 'cos')]
[('cond', <torch._ops.HigherOrderOperator object at xxx>), ('cond', <torch._ops.HigherOrderOperator object at xxx>), ('sub', <built-in function sub>)]
```
Test Plan:
See added tests in test_higher_order_ops.py and modify existing test.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/108595
Approved by: https://github.com/angelayi, https://github.com/zou3519
# Motivation
@jansel As discussed before, we expected to generalize some cuda-specific code. This can make inductor more friendly to third-party backend so that we can leverage inductor code as much as possible.
# Solution
To implement this, we give a solution to introduce device runtime abstraction. We wrapper them inside `DeviceInterface` and use `register_interface_for_device` to register each kind of device to inductor. Then use `get_interface_for_device` to fetch the corresponding runtime from device type. Then usage is like this:
```python
device_interface = get_interface_for_device("xpu")
device_interface .is_available() # to check if XPU is available
device_interface .device_count() # to check how much XPU device is available
```
The `DeviceInterface` is a simple abstraction, which enables third-party backends that implement CUDA-like semantics to be integrated with inductor. This can prevent third-party backend from using monkey patch to override some utility functions, like `decode_device` that is hard-coded with CUDA.
# Additional Context
The main code change:
- To leverage AsyncCompile, make it device-agnostic
- Avoid monkey patches, make some utility functions device-agnostic
Pull Request resolved: https://github.com/pytorch/pytorch/pull/109486
Approved by: https://github.com/jansel, https://github.com/jgong5, https://github.com/EikanWang
Summary:
Change AOTInductor to directly return output tensors instead of taking pre-allocated output tensors to return the results. This gives several benefits:
* It makes sure AOTInductor has the same behavior when managing the output tensors as the default Inductor, which is widely tested and thus more reliable.
* As we have debugged before, there are cases we still have to codegen extra copy_ ops to fill the pre-allocated output tensors which doesn't make sense for performance.
* With the coming enhanced memory planning, this again will make sure the memory planning logic is the between AOTInductor and Inductor, which will greatly simplify the problem and improve the reliability.
This change also combines D49494954 from Yang and https://github.com/pytorch/pytorch/pull/109560 from Angela.
Differential Revision: D49502318
Pull Request resolved: https://github.com/pytorch/pytorch/pull/109790
Approved by: https://github.com/chenyang78
Summary: AOT inductor is only sort-of supported on CPU right now, but it works
with a few hacks (the .so needs to be compiled and run with CUDA present,
because we haven't excised the CUDA deps; also there's an `is_cpu` flag that
needs to be plumbed into the call, or else all the weights are erroneously
allocated on GPU).
But, with those hacks in place, it currently works, so it's worth having the
current state of it continue working (and at some point we'll remove the
hacks).
Test Plan:
```
python test_aot_inductor -k test_simple_cpu
```
Reviewers: binbao
Subscribers:
Tasks:
Tags:
Differential Revision: [D49427400](https://our.internmc.facebook.com/intern/diff/D49427400)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/109625
Approved by: https://github.com/mikekgfb, https://github.com/chenyang78, https://github.com/desertfire
